High resolution network (Part 1): Principle Analysis

Computer vision can be divided into two types of tasks ： Image classification 、 Other tasks . For image classification tasks , More condensed semantic information is usually needed , High resolution characterization is not required , Just output a whole label . For target detection ( Area level classification tasks )、 Segmentation and key detection ( Pixel level classification task ) for , The network needs to be more sensitive to location information , Therefore, we need to learn high-resolution representation . How to learn image high-resolution representation is HRNet The problems solved by the network .

Traditional classification network Always From high resolution to low resolution ,HRNet It is a brand new network structure designed from scratch , It's a high resolution 、 Medium resolution and low resolution Three way parallel 、 Non stop interaction A network structure of .

It should be noted that